fread's na.strings argument should handle values like "-999"

[arunsrinivasan]

With @dselivanov's excellent PR, providing na.strings value doesn't result in columns being coerced to character anymore.

There's still one point left to address though, as mentioned under that PR -- cases like these:

require(data.table)
DT = data.table(a=9:10, b=9:10 + 0.1, c=as.logical(0:1))
text = do.call("paste", c(DT, collapse="\n", sep=","))
ans1 = fread(text, na.strings=c("9", "9.1", "FALSE"))
#    V1   V2    V3
#1:  9  9.1 FALSE
#2: 10 10.1  TRUE

sapply(ans1, class)
#         V1        V2        V3 
#  "integer" "numeric" "logical" 

# whereas read.table() gives
ans2 = read.table(text=text, na.strings=c("9", "9.1", "FALSE"), sep=",", header=FALSE)
#   V1   V2   V3
#1 NA   NA   NA
#2 10 10.1 TRUE

sapply(ans2, class)
#         V1        V2        V3 
#  "integer" "numeric" "logical" 

read.table() handles them correctly.

[arunsrinivasan]

Benchmarks on 0.1, 0.01, 0.3 and na_rich scenarios after raising the if-statement.

#    File thisFR stable (6f99651)
#     0.1 51.421 48.068
#    0.01 51.776 54.264
#   0.001 52.255 50.820
#     0.3 48.150 47.930
# na_rich 17.570 17.354

[dselivanov]

@arunsrinivasan, nice catch! Surprised by this timings!

[arunsrinivasan]

@dselivanov thanks. I wasn't clear with what the timings were. It's without na.strings=., to test default scenario. I'll add benchmarks with the arg later. I'd guess there'd be some hit. But not sure yet how to avoid that with the functionality